Research Projects
The challenge of autonomy is not replacing humans but designing systems that allow humans and machines to work together effectively.
The projects presented here were conducted during my doctoral research at the Cyber-Human-Physical Systems Lab at UC Davis, under the supervision of Dr. Zhaodan Kong. The work combined experimental platform development, neurophysiological and behavioural measurement, and computational analysis to study how humans perform and adapt in complex operational environments involving autonomous and multi-agent robotic systems. Research was supported by NASA via the HOME Space Technology Research Institute and the Air Force Office of Scientific Research.
Real-Time Inference and Prediction of Trust in Human-Autonomy Teaming
Overview
Autonomous systems that cannot detect changes in their operator’s trust are poorly equipped for real operational environments, where trust evolves as conditions change and system behavior surprises. This project developed and validated methods for continuously inferring operator trust from neurophysiological signals, without interrupting the operator or requiring explicit input.
Motivation
Human operators interacting with autonomous systems must make ongoing judgements about when to rely on the system and when to intervene. Poorly calibrated trust, whether overreliance or underuse, degrades both safety and performance. Unobtrusive real-time trust estimation opens the door to autonomous systems that can detect calibration problems as they develop and adjust their interaction strategy accordingly, rather than waiting for a breakdown to occur.
Approach
- Designed and conducted controlled experimental studies measuring EEG signals during human interaction with an autonomous system under dynamically varying trust conditions.
- Extracted spectral power features from multiple EEG channels as a baseline characterisation of neural activity.
- Computed inter-channel functional connectivity metrics capturing synchronous activity between spatially distributed brain regions as a richer representation of cognitive state during trust-relevant events.
- Compared connectivity-based features against conventional single-channel power metrics for trust prediction accuracy.
- Evaluated the relationship between neural features and dynamically changing operator trust ratings across experimental conditions.
Technical Contributions
- Demonstrated that operator trust in an autonomous system can be continuously inferred from EEG signals without requiring explicit operator input or task interruption.
- Showed that brain connectivity metrics outperform conventional single-channel power features for trust prediction, suggesting that trust is reflected in distributed neural network activity rather than localized signal changes.
- Provided a validated framework for integrating neurophysiological trust estimation into adaptive autonomy systems.
Outcomes
Network-level brain activity metrics captured trust-relevant changes more reliably than single-channel signal power, supporting the hypothesis that trust calibration involves distributed neural processes. These findings provide a technical foundation for autonomous systems capable of adjusting their behavior in response to real-time estimates of operator trust, with implications for human-robot teaming in operational environments where trust calibration directly affects mission outcomes.
Collaborators
This work is conducted in collaboration with Dr. Allie Anderson and Dr. Torin Clark at the University of Colorado Boulder.
publications
HUMAN-MultiAgent INTERACTION
Overview
Supervising a group of autonomous robots in a dynamic environment places demands on an operator that are fundamentally different from supervising a single system. The operator must maintain awareness of distributed agent behaviour, infer collective state from incomplete information, and make coordination decisions in real time. This project studied how those demands affect performance and whether the resulting operator state is detectable from neurophysiological and behavioral signals.
Motivation
Human-swarm systems are being developed for high-consequence applications including search and rescue, hazard response, and space operations, where performance breakdowns carry significant mission impact. When an operator loses track of the collective state of the robot group, coordination degrades in ways that may not be immediately visible from task metrics alone. Understanding the neurophysiological and behavioral signatures of that degradation is a necessary step toward designing systems that can detect it and respond before performance breaks down.
Approach
- Designed real-world scenario inspired experimental tasks in which participants guided a group of ground robots to sequentially assigned targets under varying conditions.
- Varied swarm configuration and task difficulty to systematically manipulate coordination demands and swarm state estimation difficulty.
- Collected gaze behavior and operator control inputs throughout task execution.
- Acquired EEG signals and computed spectral power and functional connectivity metrics to characterise operator neurophysiological state during task execution.
- Analyzed relationships between operator state measures, swarm state estimation difficulty, and task performance outcomes.
Technical Contributions
- Designed and built a human-robot interaction test facility integrating ground robot platforms, motion tracking, and multimodal physiological sensing for swarm supervision experiments.
- Identified specific performance degradation mechanisms associated with difficulty estimating the collective state of the robot group.
- Demonstrated measurable links between task difficulty, operator gaze behaviour, neurophysiological state, and performance outcomes.
- Provided an empirical basis for designing interaction policies and adaptive autonomy strategies that better support operator coordination in human-swarm teams.
Outcomes
Distinct behavioral and neurophysiological patterns were associated with different coordination strategies and levels of task performance. Operators who struggled to estimate swarm state showed measurable differences in both gaze behavior and neural activity relative to those who maintained effective awareness. These findings suggest that multimodal operator state estimation is feasible during swarm supervision tasks and could be used to trigger adaptive autonomy responses that support the operator before coordination breaks down.
publications
Digitizing Human Expertise in Manual Manufacturing
Overview
Expert manual skill in precision manufacturing is difficult to observe, harder to formalize, and currently at risk of being lost as experienced operators retire. This project developed data-driven methods for capturing and modelling the sensorimotor behaviors that distinguish expert from novice operators during precision grinding tasks, making tacit expertise visible and transferable.
Motivation
Many manufacturing outcomes still depend on skilled human operators whose knowledge lives in their hands rather than in any documented procedure. As experienced workers leave the workforce and production environments grow more complex, the inability to capture and transfer that knowledge creates real operational risk. Formalizing expert technique through data-driven analysis opens the possibility of adaptive training systems, intelligent machine tools that guide novice operators, and preservation of institutional knowledge that would otherwise disappear.
Approach
- Designed and conducted a controlled experimental study in which participants with varying levels of grinding experience performed precision manual grinding tasks.
- Collected multimodal behavioural and performance data including gaze, motion, and process outcome measurements across skill levels.
- Analyzed sensorimotor behavior patterns associated with different task execution techniques and skill levels.
- Modelled relationships between observed operator behaviours and resulting task performance outcomes.
Technical Contributions
- Developed a data-driven framework for modelling and classifying human manual expertise during precision grinding tasks.
- Identified distinguishable sensorimotor signatures associated with expert and novice operational techniques, including gaze-motor coordination patterns that predicted performance outcomes.
- Demonstrated measurable links between technique selection and task performance, providing a quantitative basis for expertise assessment.
- Established a foundation for designing manufacturing systems and training tools that adapt to operator behavior and support knowledge transfer across generations of workers.
Outcomes
Expert and novice operators showed distinct and consistently detectable differences in sensorimotor behavior during identical tasks. Gaze-motor coordination patterns in particular were strongly predictive of performance outcomes, suggesting that visual attention strategy is a core component of manual expertise that can be captured and used to guide training. These findings support the development of adaptive manufacturing systems that can assess operator skill in real time and adjust guidance or task parameters accordingly.
Collaborations
This work was conducted in collaboration with Dr. Barbara Linke at UC Davis.
